Dealing with Frequency Collapse in Time Series Embeddings by Post-Embedding reMapping
Keywords: Time Series Forcasting, deep learning, spectral analysis
Abstract: Transformer-based methods have made significant strides in time series forecasting tasks in recent years. However, we observe underfitting in numerous samples, e.g., pattern shifts or excessive deviation in extreme value regions when testing the transform-based model that converges on the training set. Through the proposed spectral analysis of adjacent embedding sequences, we identify a frequency collapse issue in the embedding features generated by the top layer of the transformer backbone. To address this, we propose the Post-Embedding ReMapping (PErM) strategy that improves the frequency-domain representation of embeddings using fixed non-linear functions. Both two kinds of PErM functions that we insert into the model can effectively resolve the frequency collapse issue and lead to significant improvements in prediction performance. Experimental results show that our method outperforms state-of-the-art algorithms across multiple datasets. We will release our code after the review phase.
Primary Area: learning on time series and dynamical systems
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 720
Loading